def test_relu(self):
top = hm.zeros((4, 12, 15, 15))
bottom = hm.random((4, 12, 15, 15), _range=(-1, 1))
@compose
def fn(bottom, top):
top = ReluForward(bottom)
return top
fn(bottom, top)
top.sync_host()
expected = np.copy(bottom)
expected[expected < 0] = 0
self._check(top, expected)
top_diff = hm.random(top.shape)
bottom_diff = hmarray(top.shape)
@compose
def fn(top_diff, bottom, bottom_diff):
bottom_diff = ReluBackward(bottom, top_diff)
return bottom_diff
fn(top_diff, bottom, bottom_diff)
bottom_diff.sync_host()
expected = np.copy(top_diff)
expected[bottom < 0] = 0
self._check(bottom_diff, expected)
def visit_Assign(self, node):
node.value = self.visit(node.value)
if not isinstance(node.targets[0], ast.Name) or node.targets[0].id in self.symbol_table:
return node
if isinstance(node.value, ast.BinOp) and len(node.targets) == 1:
if isinstance(node.value, ast.Call):
# TODO: Operations should specify an output generator
self.symbol_table[node.targets[0].id] = hm.zeros(
self.symbol_table[node.value.args[0].id].shape)
else:
if isinstance(node.value, ast.Call) and \
inspect.isclass(self.symbol_table[node.value.func.id]) and \
issubclass(self.symbol_table[node.value.func.id], HMOperation):
# TODO: Operations should specify an output generator
self.symbol_table[node.targets[0].id] = hm.zeros(
self.symbol_table[node.value.args[0].id].shape)
return node
def visit_Assign(self, node):
node.value = self.visit(node.value)
if not isinstance(node.targets[0],
ast.Name) or node.targets[0].id in self.symbol_table:
return node
if isinstance(node.value, ast.BinOp) and len(node.targets) == 1:
if isinstance(node.value, ast.Call):
# TODO: Operations should specify an output generator
self.symbol_table[node.targets[0].id] = hm.zeros(
self.symbol_table[node.value.args[0].id].shape)
else:
if isinstance(node.value, ast.Call) and \
inspect.isclass(self.symbol_table[node.value.func.id]) and \
issubclass(self.symbol_table[node.value.func.id], HMOperation):
# TODO: Operations should specify an output generator
self.symbol_table[node.targets[0].id] = hm.zeros(
self.symbol_table[node.value.args[0].id].shape)
return node
def test_pool(self):
shape = (3, 16, 24, 24)
a = hm.random(shape, _range=(0, 255))
actual_mask = hmarray((3, 16, 12, 12))
actual = hmarray((3, 16, 12, 12))
expected_mask = hmarray((3, 16, 12, 12))
expected = hmarray((3, 16, 12, 12))
expected.fill(float('-inf'))
@compose
def fn(bottom, mask, top):
top, mask = PoolForward(bottom,
kernel_size=(2, 2),
padding=(0, 0),
stride=(2, 2))
return top, mask
fn(a, actual_mask, actual)
actual.sync_host()
actual_mask.sync_host()
reference_pool(a, expected, expected_mask, (2, 2), (2, 2), (0, 0))
self._check(actual, expected)
self._check(actual_mask, expected_mask)
bottom_diff = hm.zeros(shape)
expected_bottom_diff = hm.zeros(shape)
mask = actual_mask
top_diff = hm.random((3, 16, 12, 12))
@compose
def fn(top_diff, mask, bottom_diff):
bottom_diff = PoolBackward(top_diff,
mask,
kernel_size=(2, 2),
padding=(0, 0),
stride=(2, 2))
return bottom_diff
fn(top_diff, mask, bottom_diff)
bottom_diff.sync_host()
reference_pool_backward(top_diff, mask, expected_bottom_diff, (2, 2),
(2, 2), (0, 0))
self._check(bottom_diff, expected_bottom_diff)
def test_conv_backward(self):
# TODO: Check bias diff
@compose
def fn(top_diff, weights, bottom, bottom_diff, weights_diff,
bias_diff):
bottom_diff, weights_diff, bias_diff = \
ConvBackward(bottom, top_diff, weights,
kernel_size=(11, 11), padding=(0, 0),
stride=(4, 4))
return bottom_diff, weights_diff, bias_diff
top_diff = hm.random((3, 12, 25), _range=(0, 5))
bottom = hm.random((3, 3, 27, 27), _range=(0, 255))
weights = hm.random((12, 363), _range=(-.2, .2))
weights_diff = hm.zeros((12, 363))
bias_diff = hm.zeros((12, ))
bottom_diff = hm.zeros((3, 3, 27, 27))
fn(top_diff, weights, bottom, bottom_diff, weights_diff, bias_diff)
expected_weights_diff = np.zeros_like(weights)
expected_bottom_diff = np.zeros_like(bottom_diff)
for i in range(top_diff.shape[0]):
col_data = reference_im2col(bottom[i], (11, 11), (4, 4), (0, 0))
expected_weights_diff += top_diff[i].dot(col_data.T)
expected_bottom_diff[i] = reference_col2im(
weights.T.dot(top_diff[i]), (11, 11), (4, 4), (0, 0),
expected_bottom_diff[i].shape)
weights_diff.sync_host()
np.testing.assert_array_almost_equal(weights_diff,
expected_weights_diff,
decimal=2)
bottom_diff.sync_host()
np.testing.assert_array_almost_equal(bottom_diff,
expected_bottom_diff,
decimal=2)
def test_pool(self):
shape = (3, 16, 24, 24)
a = hm.random(shape, _range=(0, 255))
actual_mask = hmarray((3, 16, 12, 12))
actual = hmarray((3, 16, 12, 12))
expected_mask = hmarray((3, 16, 12, 12))
expected = hmarray((3, 16, 12, 12))
expected.fill(float('-inf'))
@compose
def fn(bottom, mask, top):
top, mask = PoolForward(bottom, kernel_size=(2, 2),
padding=(0, 0), stride=(2, 2))
return top, mask
fn(a, actual_mask, actual)
actual.sync_host()
actual_mask.sync_host()
reference_pool(a, expected, expected_mask, (2, 2), (2, 2), (0, 0))
self._check(actual, expected)
self._check(actual_mask, expected_mask)
bottom_diff = hm.zeros(shape)
expected_bottom_diff = hm.zeros(shape)
mask = actual_mask
top_diff = hm.random((3, 16, 12, 12))
@compose
def fn(top_diff, mask, bottom_diff):
bottom_diff = PoolBackward(top_diff, mask,
kernel_size=(2, 2),
padding=(0, 0),
stride=(2, 2))
return bottom_diff
fn(top_diff, mask, bottom_diff)
bottom_diff.sync_host()
reference_pool_backward(top_diff, mask, expected_bottom_diff,
(2, 2), (2, 2), (0, 0))
self._check(bottom_diff, expected_bottom_diff)
def test_conv_backward(self):
# TODO: Check bias diff
@compose
def fn(top_diff, weights, bottom, bottom_diff, weights_diff,
bias_diff):
bottom_diff, weights_diff, bias_diff = \
ConvBackward(bottom, top_diff, weights,
kernel_size=(11, 11), padding=(0, 0),
stride=(4, 4))
return bottom_diff, weights_diff, bias_diff
top_diff = hm.random((3, 12, 25), _range=(0, 5))
bottom = hm.random((3, 3, 27, 27), _range=(0, 255))
weights = hm.random((12, 363), _range=(-.2, .2))
weights_diff = hm.zeros((12, 363))
bias_diff = hm.zeros((12, ))
bottom_diff = hm.zeros((3, 3, 27, 27))
fn(top_diff, weights, bottom, bottom_diff, weights_diff, bias_diff)
expected_weights_diff = np.zeros_like(weights)
expected_bottom_diff = np.zeros_like(bottom_diff)
for i in range(top_diff.shape[0]):
col_data = reference_im2col(bottom[i], (11, 11), (4, 4), (0, 0))
expected_weights_diff += top_diff[i].dot(col_data.T)
expected_bottom_diff[i] = reference_col2im(
weights.T.dot(top_diff[i]), (11, 11), (4, 4), (0, 0),
expected_bottom_diff[i].shape)
weights_diff.sync_host()
np.testing.assert_array_almost_equal(weights_diff,
expected_weights_diff, decimal=2)
bottom_diff.sync_host()
np.testing.assert_array_almost_equal(bottom_diff, expected_bottom_diff,
decimal=2)
def test_forward(self):
@compose
def fn(a, b, c, d):
d = ConcatForward(a, b, c)
return d
a = hm.random((16, 12, 55, 55))
b = hm.random((16, 12, 55, 55))
c = hm.random((16, 12, 55, 55))
d = hm.zeros((16, 36, 55, 55))
d = fn(a, b, c, d)
d.sync_host()
np.testing.assert_array_almost_equal(d[:16, 0:12, ...], a)
np.testing.assert_array_almost_equal(d[:16, 12:24, ...], b)
np.testing.assert_array_almost_equal(d[:16, 24:36, ...], c)
def test_conv_forward(self):
@compose
def fn(a, weights, out, bias):
out = ConvForward(a, weights, bias, kernel_size=(11, 11),
padding=(0, 0), stride=(4, 4))
return out
a = hm.random((3, 3, 27, 27), _range=(0, 255))
weights = hm.random((12, 363), _range=(-.2, .2))
out = hm.zeros((3, 12, 5, 5))
bias = hm.random((12, ))
fn(a, weights, out, bias)
weights = weights.reshape(12, 3, 11, 11)
expected = np.zeros((3, 12, 5, 5), np.float32)
reference_conv(a, weights, bias, expected, (4, 4), (0, 0))
out.sync_host()
np.testing.assert_array_almost_equal(out, expected,
decimal=1)
def test_conv_forward(self):
@compose
def fn(a, weights, out, bias):
out = ConvForward(a,
weights,
bias,
kernel_size=(11, 11),
padding=(0, 0),
stride=(4, 4))
return out
a = hm.random((3, 3, 27, 27), _range=(0, 255))
weights = hm.random((12, 363), _range=(-.2, .2))
out = hm.zeros((3, 12, 5, 5))
bias = hm.random((12, ))
fn(a, weights, out, bias)
weights = weights.reshape(12, 3, 11, 11)
expected = np.zeros((3, 12, 5, 5), np.float32)
reference_conv(a, weights, bias, expected, (4, 4), (0, 0))
out.sync_host()
np.testing.assert_array_almost_equal(out, expected, decimal=1)
from hindemith.core import compose import caffe import numpy as np import time prototxt = "models/alexnet-ng/deploy.prototxt" caffemodel = "models/alexnet-ng/alexnet-ng.caffemodel" # caffe.set_mode_gpu() # caffe.set_device(2) # caffe.set_mode_cpu() caffe_net = caffe.Net(prototxt, caffemodel, caffe.TEST) conv1_filters = caffe_net.params['conv1'][0].data.view(hmarray) conv1_bias = caffe_net.params['conv1'][1].data.view(hmarray) conv1 = hm.zeros(caffe_net.blobs['conv1'].data.shape) norm1 = hm.zeros(caffe_net.blobs['norm1'].data.shape) norm1_scale = hm.zeros(norm1.shape) pool1 = hm.zeros(caffe_net.blobs['pool1'].data.shape) pool1_mask = hm.zeros(pool1.shape) conv2_filters = caffe_net.params['conv2'][0].data.view(hmarray) conv2_bias = caffe_net.params['conv2'][1].data.view(hmarray) conv2 = hm.zeros(caffe_net.blobs['conv2'].data.shape) norm2 = hm.zeros(caffe_net.blobs['norm2'].data.shape) norm2_scale = hm.zeros(norm2.shape) pool2 = hm.zeros(caffe_net.blobs['pool2'].data.shape)
import pycl as cl import caffe import numpy as np import time prototxt = "benchmarks/alexnet.prototxt" caffemodel = "models/alexnet-ng/alexnet-ng.caffemodel" caffe.set_mode_gpu() caffe.set_device(2) # caffe.set_mode_cpu() caffe_net = caffe.Net(prototxt, caffemodel, caffe.TEST) conv1_filters = caffe_net.params['conv1'][0].data.view(hmarray) conv1_bias = caffe_net.params['conv1'][1].data.view(hmarray) conv1 = hm.zeros(caffe_net.blobs['conv1'].data.shape) norm1 = hm.zeros(caffe_net.blobs['norm1'].data.shape) norm1_scale = hm.zeros(norm1.shape) pool1 = hm.zeros(caffe_net.blobs['pool1'].data.shape) pool1_mask = hm.zeros(pool1.shape) conv2_filters = caffe_net.params['conv2'][0].data.view(hmarray) conv2_bias = caffe_net.params['conv2'][1].data.view(hmarray) conv2 = hm.zeros(caffe_net.blobs['conv2'].data.shape) norm2 = hm.zeros(caffe_net.blobs['norm2'].data.shape) norm2_scale = hm.zeros(norm2.shape) pool2 = hm.zeros(caffe_net.blobs['pool2'].data.shape)
import pycl as cl import caffe import numpy as np import time prototxt = "benchmarks/alexnet.prototxt" caffemodel = "models/alexnet-ng/alexnet-ng.caffemodel" caffe.set_mode_gpu() caffe.set_device(2) # caffe.set_mode_cpu() caffe_net = caffe.Net(prototxt, caffemodel, caffe.TEST) conv1_filters = caffe_net.params["conv1"][0].data.view(hmarray) conv1_bias = caffe_net.params["conv1"][1].data.view(hmarray) conv1 = hm.zeros(caffe_net.blobs["conv1"].data.shape) norm1 = hm.zeros(caffe_net.blobs["norm1"].data.shape) norm1_scale = hm.zeros(norm1.shape) pool1 = hm.zeros(caffe_net.blobs["pool1"].data.shape) pool1_mask = hm.zeros(pool1.shape) conv2_filters = caffe_net.params["conv2"][0].data.view(hmarray) conv2_bias = caffe_net.params["conv2"][1].data.view(hmarray) conv2 = hm.zeros(caffe_net.blobs["conv2"].data.shape) norm2 = hm.zeros(caffe_net.blobs["norm2"].data.shape) norm2_scale = hm.zeros(norm2.shape) pool2 = hm.zeros(caffe_net.blobs["pool2"].data.shape)